Update README.md

README.md

@@ -71,8 +71,12 @@ when batch_size is large (16, 32, 64, 128), throughput_amp > throughput_tf32.
 
 - Observation 2: The coefficient of variation of throughput for the 100 iterations is smallest when batch_size = 128.  
 
-Benchmarking with batch_size = 128 can be used for node health check. For example, the expected throughput for dim = 2, nodes = 1, gpus = 4, batch_size = 128, tf32 would be 2500 ± 200.
+Benchmarking with batch_size = 128 can be used for node health check. For example, the expected throughput for dim = 2, nodes = 1, gpus = 4, batch_size = 128 would be 2500 ± 200 (TF32) and 4100 ± 400 (AMP).
 
 <img src="https://github.com/xuagu37/Benchmark_nnU-Net_for_PyTorch/blob/main/figures/benchmark_cv_example.png" width="400">
 
+- Observation 3: Ideally, throughput would be douled when batch_size doubled. In practice, throughtput does not increase ideally when batch_size > 16.
+
+<img src="https://github.com/xuagu37/Benchmark_nnU-Net_for_PyTorch/blob/main/figures/benchmark_batch_size_ideal.png" width="400">
+